nir: Add a fused multiply-add peephole

src/glsl/Makefile.sources

@@ -32,6 +32,7 @@ NIR_FILES = \
 	$(GLSL_SRCDIR)/nir/nir_opt_copy_propagate.c \
 	$(GLSL_SRCDIR)/nir/nir_opt_dce.c \
 	$(GLSL_SRCDIR)/nir/nir_opt_global_to_local.c \
+	$(GLSL_SRCDIR)/nir/nir_opt_peephole_ffma.c \
 	$(GLSL_SRCDIR)/nir/nir_opt_peephole_select.c \
 	$(GLSL_SRCDIR)/nir/nir_print.c \
 	$(GLSL_SRCDIR)/nir/nir_remove_dead_variables.c \

src/glsl/nir/nir.h

@@ -1374,6 +1374,7 @@ bool nir_opt_dce_impl(nir_function_impl *impl);
 bool nir_opt_dce(nir_shader *shader);
 
 bool nir_opt_peephole_select(nir_shader *shader);
+bool nir_opt_peephole_ffma(nir_shader *shader);
 
 #ifdef __cplusplus
 } /* extern "C" */

src/glsl/nir/nir_opcodes.h

@@ -307,6 +307,7 @@ BINOP_HORIZ(vec2, 2, nir_type_unsigned, 1, nir_type_unsigned, 1, nir_type_unsign
    ARR(src1_size, src2_size, src3_size), \
    ARR(nir_type_unsigned, nir_type_unsigned, nir_type_unsigned))
 
+/* fma(a, b, c) = (a * b) + c */
 TRIOP(ffma, nir_type_float)
 
 TRIOP(flrp, nir_type_float)

src/glsl/nir/nir_opt_peephole_ffma.c

@@ -0,0 +1,191 @@
+/*
+ * Copyright © 2014 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ * Authors:
+ *    Jason Ekstrand (jason@jlekstrand.net)
+ *
+ */
+
+#include "nir.h"
+
+/*
+ * Implements a small peephole optimization that looks for a multiply that
+ * is only ever used in an add and replaces both with an fma.
+ */
+
+struct peephole_ffma_state {
+   void *mem_ctx;
+   nir_function_impl *impl;
+   bool progress;
+};
+
+static inline nir_alu_instr *
+get_mul_for_src(nir_alu_instr *add, unsigned idx)
+{
+   if (!add->src[idx].src.is_ssa)
+      return NULL;
+
+   /* We can't handle these in between the operations */
+   if (add->src[idx].negate || add->src[idx].abs)
+      return NULL;
+
+   nir_instr *instr = add->src[idx].src.ssa->parent_instr;
+   if (instr->type != nir_instr_type_alu)
+      return NULL;
+
+   nir_alu_instr *mul = nir_instr_as_alu(instr);
+   if (mul->op != nir_op_fmul)
+      return NULL;
+
+   /* Can't handle a saturate in between */
+   if (mul->dest.saturate)
+      return NULL;
+
+   /* We already know that the same source is not used twice in the add and
+    * we will assume valid use-def information, so this check is sufficient
+    */
+   if (mul->dest.dest.ssa.uses->entries > 1)
+      return NULL; /* Not the only use */
+
+   return mul;
+}
+
+/* Copies (and maybe swizzles) the given ALU source */
+static inline void
+copy_alu_src(void *mem_ctx, nir_alu_src *new_src, nir_alu_src old_src,
+             uint8_t *swizzle)
+{
+   new_src->src = nir_src_copy(old_src.src, mem_ctx);
+   new_src->abs = old_src.abs;
+   new_src->negate = old_src.negate;
+
+   if (swizzle == NULL) {
+      memcpy(new_src->swizzle, old_src.swizzle, sizeof old_src.swizzle);
+   } else {
+      for (int i = 0; i < 4; ++i) {
+         if (swizzle[i] < 4)
+            new_src->swizzle[i] = old_src.swizzle[swizzle[i]];
+      }
+   }
+}
+
+static bool
+nir_opt_peephole_ffma_block(nir_block *block, void *void_state)
+{
+   struct peephole_ffma_state *state = void_state;
+
+   nir_foreach_instr_safe(block, instr) {
+      if (instr->type != nir_instr_type_alu)
+         continue;
+
+      nir_alu_instr *add = nir_instr_as_alu(instr);
+      if (add->op != nir_op_fadd)
+         continue;
+
+      /* TODO: Maybe bail if this expression is considered "precise"? */
+
+      /* This, is the case a + a.  We would rather handle this with an
+       * algebraic reduction than fuse it.  Also, we want to only fuse
+       * things where the multiply is used only once and, in this case,
+       * it would be used twice by the same instruction.
+       */
+      if (add->src[0].src.is_ssa && add->src[1].src.is_ssa &&
+          add->src[0].src.ssa == add->src[1].src.ssa)
+         continue;
+
+      nir_alu_instr *mul = get_mul_for_src(add, 0);
+      unsigned mul_src = 0;
+
+      if (mul == NULL) {
+         mul = get_mul_for_src(add, 1);
+         mul_src = 1;
+      }
+
+      if (mul == NULL)
+         continue;
+
+      nir_alu_instr *ffma = nir_alu_instr_create(state->mem_ctx, nir_op_ffma);
+      ffma->dest.saturate = add->dest.saturate;
+      ffma->dest.write_mask = add->dest.write_mask;
+
+      copy_alu_src(state->mem_ctx, &ffma->src[0], mul->src[0],
+                   add->src[mul_src].swizzle);
+      copy_alu_src(state->mem_ctx, &ffma->src[1], mul->src[1],
+                   add->src[mul_src].swizzle);
+      copy_alu_src(state->mem_ctx, &ffma->src[2], add->src[1 - mul_src], NULL);
+
+      if (add->dest.dest.is_ssa) {
+         ffma->dest.dest.is_ssa = true;
+         nir_ssa_def_init(state->impl, &ffma->instr, &ffma->dest.dest.ssa,
+                          add->dest.dest.ssa.num_components,
+                          add->dest.dest.ssa.name);
+
+         nir_src ffma_dest_src = {
+            .is_ssa = true,
+            .ssa = &ffma->dest.dest.ssa,
+         };
+         nir_ssa_def_rewrite_uses(&add->dest.dest.ssa, ffma_dest_src,
+                                  state->mem_ctx);
+      } else {
+         ffma->dest.dest = nir_dest_copy(add->dest.dest, state->mem_ctx);
+      }
+
+      nir_instr_insert_before(&add->instr, &ffma->instr);
+      nir_instr_remove(&add->instr);
+      nir_instr_remove(&mul->instr);
+
+      state->progress = true;
+   }
+
+   return true;
+}
+
+static bool
+nir_opt_peephole_ffma_impl(nir_function_impl *impl)
+{
+   struct peephole_ffma_state state;
+
+   state.mem_ctx = ralloc_parent(impl);
+   state.impl = impl;
+   state.progress = false;
+
+   nir_foreach_block(impl, nir_opt_peephole_ffma_block, &state);
+
+   if (state.progress)
+      nir_metadata_dirty(impl, nir_metadata_block_index |
+                               nir_metadata_dominance);
+
+   return state.progress;
+}
+
+bool
+nir_opt_peephole_ffma(nir_shader *shader)
+{
+   bool progress = false;
+
+   nir_foreach_overload(shader, overload) {
+      if (overload->impl)
+         progress |= nir_opt_peephole_ffma_impl(overload->impl);
+   }
+
+   return progress;
+}

src/mesa/drivers/dri/i965/brw_fs_nir.cpp

@@ -63,6 +63,8 @@ fs_visitor::emit_nir_code()
       nir_validate_shader(nir);
       nir_opt_peephole_select(nir);
       nir_validate_shader(nir);
+      nir_opt_peephole_ffma(nir);
+      nir_validate_shader(nir);
       nir_print_shader(nir, stderr);
 
       nir_convert_from_ssa(nir);